Re: bpf: add support to check kernel features in BPF program

[zf <zf15750701@gmail.com>]

在 2023/5/4 19:09, Menglong Dong 写道:
> Hello,
> 
> I find that it's not supported yet to check if the bpf features are
> supported by the target kernel in the BPF program, which makes
> it hard to keep the BPF program compatible with different kernel
> versions.
> 
> For example, I want to use the helper bpf_jiffies64(), but I am not
> sure if it is supported by the target, as my program can run in
> kernel 5.4 or kernel 5.10. Therefore, I have to compile two versions
> BPF elf and load one of them according to the current kernel version.
> The part of BPF program can be this:
> 
> #ifdef BPF_FEATS_JIFFIES64
>    jiffies = bpf_jiffies64();
> #else
>    jiffies = 0;
> #endif
> 
> And I will generate xxx_no_jiffies.skel.h and xxx_jiffies.skel.h
> with -DBPF_FEATS_JIFFIES64 or not.
> 
> This method is too silly, as I have to compile 8(2*2*2) versions of
> the BPF program if I am not sure if 3 bpf helpers are supported by the
> target kernel.
> 
> Therefore, I think it may be helpful if we can check if the helpers
> are support like this:
> 
> if (bpf_core_helper_exist(bpf_jiffies64))
>    jiffies = bpf_jiffies64();
> else
>    jiffies = 0;
> 
> And bpf_core_helper_exist() can be defined like this:
> 
> #define bpf_core_helper_exist(helper)                        \
>      __builtin_preserve_helper_info(helper, BPF_HELPER_EXISTS)
> 
> Besides, in order to prevent the verifier from checking the helper
> that is not supported, we need to remove the dead code in libbpf.
> As the kernel already has the ability to remove dead and nop insn,
> we can just make the dead insn to nop.
> 
> Another option is to make the BPF program support "const value".
> Such const values can be rewrite before load, the dead code can
> be removed. For example:
> 
> #define bpf_const_value __attribute__((preserve_const_value))
> 
> bpf_const_value bool is_bpf_jiffies64_supported = 0;
> 
> if (is_bpf_jiffies64_supported)
>    jiffies = bpf_jiffies64();
> else
>    jiffies = 0;
> 
> The 'is_bpf_jiffies64_supported' will be compiled to an imm, and
> can be rewrite and relocated through libbpf by the user. Then, we
> can make the dead insn 'nop'.
> 
> What do you think? I'm not sure if these methods work and want
> to get some advice before coding.
> 
> Thanks!
> Menglong Dong

Maybe you can use like this
bpf_core_enum_value_exists(enum bpf_func_id, BPF_FUNC_jiffies64);